Content area
The increasing complexity of modern wireless networks demands predictive strategies for proactive adaptation and sustained Quality of Service. This thesis presents a multi-horizon predictive analytics framework that delivers actionable insights for dynamic wireless network management. The framework consists of two complementary components. First, a Random Forest-based signal trend classifier forecasts the directional change of signal strength (up, down, or flat) over the next 1 to 5 seconds, using current signal values, spatial data, and engineered temporal features such as rolling mean and signal variation. Second, a multi-step packet loss classifier predicts cumulative packet loss classes across the same horizons using network features including signal strength, distances, and antenna configuration. To bridge the gap between raw network metrics and application-level outcomes, the framework incorporates a task-aware mapping that links each predicted loss class to performance degradation of downstream applications, such as image classification (e.g., VGG16, ResNet9) and instance segmentation (e.g., YOLOv3, YOLOv8). Decision confidence is also quantified to support risk-aware adaptation. In addition to forecasting, the framework serves as a semantic prediction module for task-aware control systems such as SOAR. By predicting both network degradation and its expected impact on task performance, the system can guide intelligent configuration decisions. These include adjusting MU-MIMO stream count, modifying packet duplication strategies, or scheduling tasks more efficiently. Importantly, when predicted task performance remains within acceptable limits, even under moderate packet loss, the system may choose not to reconfigure. This prevents unnecessary resource consumption and promotes more efficient operation in real-time edge computing environments. Experimental results demonstrate high accuracy across all forecasting horizons, validating the framework’s ability to predict network dynamics. By combining predictive precision, semantic relevance, and interpretability, the proposed framework provides a practical foundation for intelligent and task-sensitive network control in wireless edge environments.